Pashto Handwritten Invariant Character Trajectory Prediction Using a Customized Deep Learning Technique.

Area of Science:

• Computer Science
• Artificial Intelligence
• Natural Language Processing

Background:

• Historical documents and communication relied heavily on handwritten records across diverse ethnic groups.
• Automation efforts for handwritten text recognition primarily focused on major languages, neglecting regional and minor languages.
• Pashto, a cursive script vital to Pakistan and Afghanistan, requires preservation through automated recognition.

Purpose of the Study:

• To develop and evaluate deep learning models for the detection and recognition of Pashto handwritten characters and ligatures.
• To enhance a Pashto handwritten dataset through data augmentation and noise reduction techniques.
• To compare the performance of various deep learning architectures for Pashto script recognition.

Main Methods:

• A new dataset of Pashto handwritten characters and ligatures was created and augmented using scaling and rotation.
• Morphological operations and median filtering were applied for data preprocessing and noise removal.
• Several deep learning models, including VGG19, MobileNetV2, MobileNetV3, and a customized Convolutional Neural Network (CNN), were trained and evaluated.

Main Results:

• The customized CNN model achieved the highest accuracy and lowest loss among the evaluated deep learning techniques.
• The customized CNN model attained a training accuracy of 93.98%, validation accuracy of 92.08%, and testing accuracy of 92.99%.
• Data augmentation and preprocessing techniques significantly improved the robustness of the recognition models.

Conclusions:

• Deep learning, particularly a customized CNN, is effective for recognizing Pashto handwritten characters and ligatures.
• The developed methods contribute to the preservation and accessibility of the Pashto language.
• This research highlights the potential for applying advanced AI techniques to under-resourced languages.